Watching Jensen’s lecture was cool. Some things i noticed: Jensen talks for 99% of the lecture, in past lectures Anjey usually steered the conversation. He insisted on saying what he wanted to say over taking questions.

The old days

Computer Design stayed the same for 64 years

• but now computers respond + take initiative

So in the past you had people working on specialized topics.

New: Codesigning

But now to make good chips, people must work together

• Ex. you need microprocessor engineers and compiler engineers working together

You can even outperform Moores Law with codesign

are the old days useless?

• Although the first principles you learn are exhausted in processor deisng, they are still IMPORTANT

how can open source models improve MFU

When asked how the open-source ecosystem can improve Model Flops Utilization (MFU) without being a vertically integrated giant that controls the entire hardware and software stack, Jensen flipped the premise of the question entirely.

HIGH MFU IS NOT THE GOAL (MFU is % of GPU USAGE)

• Id prefer a model that does better in less compute

Overprovisioning to fight Amdahl’s Law

• Large data centers handle shifting bottlenecks across raw flops, memory bandwidth, memory capacity, and network capacity.
• To avoid being stalled by a single bottleneck (Amdahl’s Law), you want to overprovision your compute. Flops are functionally cheap; what you are actually paying for is bandwidth and architecture

Instead, focus on real-world evaluations—such as maximizing intelligence delivered per watt—and build architectures that balance multi-domain capability without over-fitting to a single problem

• So having good ‘taste’ is valuable - enjoyment , trial and error, what should we not do , etc. ITS AN ART

the evolutionary line of NVIDIA’s chips

each generation of NVIDIA solves a bottleneck from first princples Hopper -> Blackwell-> Vera Rubin-> Feynman (Pre-training) (LLM Inference) (AI Agents) (Swarm of Agents)

Hopper is built for training massive foundation models

• told it was a bad idea to build a system nobody could afford
• but it paid off

Blackwell

• LLM decoding required massive bandwidth, so they put 72 chips (NVLink 72) to act as a GPU together

Vera Rubin

• AI agents rely on context + memory, but copying over the network is too slow.

• Rubin connect storage directly to the processor. when traditional GPU invokes an AI agent, it is stuck sitting idle, waiting for that CPU to respond.

• Vera CPU: Instead of hundreds of slow parallel cores, it focuses on providing ultra-low-latency, high-performance single-threaded execution per core so tools return results to the GPU instantly

Feynman

• we transition to swarms of agents, controlled by Feynman

Jensens’ intuition Framework

Observe the Shift: Watch real-world deviations (e.g., AlexNet crushing traditional computer vision)

Deconstruct to First Principles: Strip away legacy assumptions (e.g., “CPUs must be built for traditional cloud servers”)

Identify the Computing Pattern: Map out how the new software behaves—does it need raw computing power, extreme memory bandwidth, or hyper-low single-threaded latency?

Build the Mental Model: - What am i observing? Reason back to first principle - Whats going to happen next? What else can it solve? What does it mean to computers - Keep iteration until first principles until you get a mental model on computing for a domain !

Work Backwards to System Design: Build the specialized chips, fabrics, and software stacks to support that future completely.

college advice

What should students focus on?

• don’t focus on your passion
• If you have a passion sure, but finding a passion is too high a bar
• Just do whatever you’re good at

Condition yourself to always look to get somewhere better no matter how hard

• Ie. suffering teaches u resilience
• Ur family, and job needs u to be resilient , u need to work that muscle
• Dont seek for just joy but suffering and pain
• 90% of my job is hard, and i do not enjoy (Jensen)

importance of democratizing Chips

Everyone already has NVIDIA chips, there is nothing wrong with selling globally

people who say we don’t know how AI works is a liar

Shouldn’t people in US get first access to CHIPS? - NO

• but if the dean of Stanford orders a bunch of chips, I’ll deliver

If someone orders chips we DELIVER 100%. BUT THE ISSUE IS: THE SYSTEM IS NOT READY TO DELIVER COMPUTE

For example, at universities everyone gets fundings but gets compute only for themselves.

• We should aggregate compute and have a supercomputer that everyone shares by the cloud. It takes a large investment which nobody is willing to make.